| content |
{"Commits":[{"Sha1":"8b0729c9b {"Commits":[{"Sha1":"8b0729c9b6d46b1dd46a9f2f4b7536a5106a812e","Message":"fix(engine): Tesla P4 出歌链路修复(fp32+int8量化+DiT CPU offload)\n\nP4(sm_61 Pascal, 8GB) 上出歌此前必失败,根因是四个耦合约束:\n- fp16(默认)在 DiT 采样溢出产 NaN latents;唯一数值稳定路径是 fp32\n- fp32 DiT 权重 ~7.3GB,引擎「先整体搬 GPU 再量化」,加载即 OOM\n- 引擎 HTTP 出歌懒加载路径漏传 quantization,tier3 的量化默认成死代码\n- Pascal 默认降级的 w8a8_dynamic 张量不支持跨设备搬运,与 DiT offload 冲突\n\n有效组合: fp32 + int8_weight_only 量化 + DiT CPU offload + expandable_segments,\nDiT 在 CPU 上量化为 int8 后分层搬 GPU,常驻显存仅 ~387MB。实测 30s 歌成功落盘 FLAC。\n\n- start-engine-p4.sh: 注入上述全部运行时开关\n- deploy/ace-step/patches/: 固化引擎侧改动(startup_model_init.py 接通 quantization)\n- bootstrap-engine.sh / Dockerfile.tesla-p4: 引擎重建时自动应用补丁(引擎目录 gitignored)\n","AuthorEmail":"luoguoguo@gmail.com","AuthorName":"luoguoguo","CommitterEmail":"luoguoguo@gmail.com","CommitterName":"luoguoguo","Timestamp":"2026-07-30T12:42:31+08:00"},{"Sha1":"c53b6955d7851b95d797b90087813d6a2a83d32f","Message":"feat: 新增 Tesla P4 引擎启动脚本(禁用 cuDNN)\n\n随 torch 2.5.1+cu121 的 cuDNN 9.1.9 在 Pascal(sm_61) 上找不到卷积算法\n（DiT proj_in Conv1d 报 \"GET was unable to find an engine to execute this\ncomputation\"，fp16/fp32 皆然）。P4 无 Tensor Core，cuDNN 不带来加速，\n故启动时注入 torch.backends.cudnn.enabled=False，卷积回退原生 CUDA kernel。\n同时套用 Pascal 兼容开关（5Hz LM/LLM 走 pt backend）。\n","AuthorEmail":"luoguoguo@gmail.com","AuthorName":"luoguoguo","CommitterEmail":"luoguoguo@gmail.com","CommitterName":"luoguoguo","Timestamp":"2026-07-30T09:40:48+08:00"},{"Sha1":"51cd9fe059767d2377bd6292cf1913a9885f8cd7","Message":"feat: 歌词起草接入豆包大模型 + 本地/公网运行配置\n\n歌词起草从纯前端模板改为调用豆包（火山引擎 Ark，OpenAI 兼容）大模型：\n- contracts 新增 lyrics 域（draftLyricSchema 入参/响应契约）\n- 后端新建 lyrics 模块：POST /api/lyrics/draft，未配置 DOUBAO_API_KEY 时 503\n- 前端起草按钮改调后端，大模型不可用时降级回本地模板草稿，不阻断创作\n- .env.example 登记 DOUBAO_API_KEY/BASE_URL/MODEL（真实密钥仅在 gitignored 的 .env）\n\n运行配置：docker-compose Redis 端口参数化、前端 dev 绑 0.0.0.0 以支持公网访问\n","AuthorEmail":"luoguoguo@gmail.com","AuthorName":"luoguoguo","CommitterEmail":"luoguoguo@gmail.com","CommitterName":"luoguoguo","Timestamp":"2026-07-29T18:01:05+08:00"}],"HeadCommit":{"Sha1":"8b0729c9b6d46b1dd46a9f2f4b7536a5106a812e","Message":"fix(engine): Tesla P4 出歌链路修复(fp32+int8量化+DiT CPU offload)\n\nP4(sm_61 Pascal, 8GB) 上出歌此前必失败,根因是四个耦合约束:\n- fp16(默认)在 DiT 采样溢出产 NaN latents;唯一数值稳定路径是 fp32\n- fp32 DiT 权重 ~7.3GB,引擎「先整体搬 GPU 再量化」,加载即 OOM\n- 引擎 HTTP 出歌懒加载路径漏传 quantization,tier3 的量化默认成死代码\n- Pascal 默认降级的 w8a8_dynamic 张量不支持跨设备搬运,与 DiT offload 冲突\n\n有效组合: fp32 + int8_weight_only 量化 + DiT CPU offload + expandable_segments,\nDiT 在 CPU 上量化为 int8 后分层搬 GPU,常驻显存仅 ~387MB。实测 30s 歌成功落盘 FLAC。\n\n- start-engine-p4.sh: 注入上述全部运行时开关\n- deploy/ace-step/patches/: 固化引擎侧改动(startup_model_init.py 接通 quantization)\n- bootstrap-engine.sh / Dockerfile.tesla-p4: 引擎重建时自动应用补丁(引擎目录 gitignored)\n","AuthorEmail":"luoguoguo@gmail.com","AuthorName":"luoguoguo","CommitterEmail":"luoguoguo@gmail.com","CommitterName":"luoguoguo","Timestamp":"2026-07-30T12:42:31+08:00"},"CompareURL":"luoanwu/song-generation/compare/d28a8594e6cf9091cdf9ccd0fad377041dfaf727...8b0729c9b6d46b1dd46a9f2f4b7536a5106a812e","Len":3}... |